Method, device, readable storage medium and equipment for silent face liveness detection

Abstract

The present invention discloses a method, device, readable storage medium and equipment for silent liveness detection of a face, which belongs to the field of face recognition. The method comprises: inputting an image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is the probability value of a live object or a prosthesis in the receptive field of the corresponding position of the image to be detected; statistically analyzing the feature values ​​of all positions on the target feature map to obtain a liveness detection score of the image to be detected; comparing the liveness detection score of the image to be detected with a set threshold, and if the set threshold is met, it is judged to be alive, otherwise it is judged to be a prosthesis. Wherein, when the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing a live object or a prosthesis. The present invention has good robustness, reduces equipment cost, makes full use of local detail information of the face, and improves the accuracy of liveness judgment.

Description

Technical Field

[0001] The present invention relates to the field of face recognition, and in particular to a method, device, readable storage medium and equipment for silent face liveness detection. Background Art

[0002] In recent years, facial recognition, owing to its convenience and contactlessness, has gained widespread popularity and application in finance, security, and other fields. However, its ease of acquisition also makes it easy for others to exploit it, creating fake faces by printing photos or shooting videos, attacking facial recognition systems and impersonating others. Therefore, liveness detection based on facial images is extremely important and a prerequisite for facial recognition.

[0003] Currently, there are two common types of facial liveness detection technologies: user-coordinated liveness detection and silent liveness detection. User-coordinated liveness detection involves the liveness detection system randomly sending multiple facial movement commands, requiring the user to respond sequentially within a specified timeframe. The system then determines whether the face is live based on the accuracy of the user's responses. Silent liveness detection involves the user silently completing the liveness detection task without any awareness. The two definitions of liveness detection clearly demonstrate that silent liveness detection breaks the "completing specified movements = I'm alive" formula of cooperative liveness detection. It requires no user cooperation, is more user-friendly, and offers fast detection speeds and a superior user experience.

[0004] The currently used silent face liveness detection methods mainly include traditional statistical learning-based methods, 3D-based methods, and near-infrared-based methods.

[0005] (1) Silent face detection based on traditional statistical learning methods

[0006] Traditional statistical learning-based silent face liveness detection methods use hand-crafted features to extract features from images. These features are then used for binary classification to determine whether the image is live or not. The most common feature used in this method is the LBP feature. This method uses the LBP operator to extract features from facial images, then segments the encoded features into blocks and generates statistical feature maps. Finally, support vector machines (SVMs) are used for binary classification. This method relies on the quality of hand-crafted features, making it less robust to variations in posture, lighting, and other factors.

[0007] (2) 3D-based silent face detection

[0008] This method uses 3D structured light to image the human face. Live human faces have clear texture and depth information under 3D structured light, making them suitable for face liveness detection. However, prosthetic images have unclear or no texture under 3D structured light, and lack depth information. While this method is accurate for silent liveness detection, it places high demands on equipment and is expensive.

[0009] (3) Silent face detection method based on near-infrared

[0010] Near-infrared silent face liveness detection relies primarily on the differences in how different images appear when viewed under near-infrared light. Near-infrared imaging has the following characteristics: photos, videos, and inkjet prints displayed on screens do not produce near-infrared images. Printed photos, masks, and other items exhibit different reflectivity in near-infrared light than real people. This is primarily reflected in the following ways: Real people's eyes and noses produce light spots, while printed photos generally do not; real people's eyes appear whitish, while printed photos generally do not; and real people's cheeks appear dark, while printed photos generally appear bright due to strong light reflections. This method offers high accuracy, but requires a near-infrared light source and has high requirements for ambient lighting.

[0011] It should also be pointed out that the existing technology directly uses face images for binary classification tasks, which is prone to overfitting. When training the convolutional neural network, the entire training sample is directly labeled as 0 or 1. For example, the label of the real training sample is set to 1, and the label of the prosthesis training sample is set to 0. The convolutional neural network is trained based on the training sample, and then the face image to be tested is input into the trained convolutional neural network. It only uses the global information of the face image for liveness detection, which is equivalent to directly classifying an entire area and cannot fully utilize the detailed information of the face image. Therefore, the existing liveness detection method has poor classification effect and low accuracy. Summary of the Invention

[0012] In order to address the defects of the existing silent face liveness detection methods in terms of poor robustness, high cost, and poor liveness detection effect, the present invention provides a silent face liveness detection method, device, readable storage medium and equipment. The present invention has good robustness, reduces equipment cost, fully utilizes the local detail information of the face, and improves the accuracy of liveness judgment.

[0013] The present invention provides the following technical solutions:

[0014] In a first aspect, the present invention provides a method for detecting silent liveness of a face, the method comprising:

[0015] Input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is the probability value of the receptive field of the corresponding position of the image to be detected being a living object or a prosthesis;

[0016] Perform statistical analysis on the feature values ​​of all positions on the target feature map to obtain the liveness detection score of the image to be detected;

[0017] Compare the liveness detection score of the image to be detected with the threshold set. If it meets the set threshold, it is judged as a live body, otherwise it is judged as a fake body;

[0018] When the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing a living body or a prosthesis.

[0019] In a second aspect, the present invention provides a silent face liveness detection device, the device comprising:

[0020] A processing module is configured to input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is a probability value of a living object or a prosthesis in the receptive field of the corresponding position of the image to be detected;

[0021] The statistical module is used to perform statistical analysis on the feature values ​​of all positions on the target feature map to obtain the liveness detection score of the image to be detected;

[0022] A judgment module is used to compare the liveness detection score of the image to be detected with a set threshold. If the set threshold is met, it is judged to be alive, otherwise it is judged to be a prosthesis;

[0023] When the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing a living body or a prosthesis.

[0024] In a third aspect, the present invention provides a computer-readable storage medium for silent face liveness detection, comprising a memory for storing processor-executable instructions, which, when executed by the processor, implements the steps of the silent face liveness detection method described in the first aspect.

[0025] In a fourth aspect, the present invention provides a device for silent face liveness detection, comprising at least one processor and a memory storing computer-executable instructions, wherein when the processor executes the instructions, the steps of the silent face liveness detection method described in the first aspect are implemented.

[0026] The present invention has the following beneficial effects:

[0027] 1. Use convolutional neural networks for liveness detection, which has good robustness. It can also detect liveness in visible light facial images, which can reduce equipment costs and has low requirements for lighting environments.

[0028] 2. Each pixel in the training sample is marked with a live or prosthetic label value, and the live and prosthetic images are annotated point by point for liveness judgment. This can not only improve the accuracy of liveness judgment, but also effectively reflect which areas have the highest impact on liveness detection.

[0029] 3. Operate the target feature map obtained by CNN on the image to be detected, and judge the liveness judgment result on a smaller receptive field of the image to be detected. No longer use softmax to reduce the possibility of overfitting. At the same time, make full use of the local detail information of the face to make the liveness judgment more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flowchart of one embodiment of the method for silent face liveness detection of the present invention;

[0031] Figure 2 Flowchart of the training process of the convolutional neural network of the present invention;

[0032] Figure 3 for Figure 2 The flowchart of S100' in the flowchart of the training process of the convolutional neural network;

[0033] Figure 4 An example diagram for key point localization of real / prosthetic face images and real / prosthetic sub-region division;

[0034] Figure 5 Schematic diagram of the correspondence between the position of the target feature map and the receptive field of the image to be detected;

[0035] Figure 6 This is a flowchart of another embodiment of the method for silent face liveness detection of the present invention;

[0036] Figure 7 Schematic diagram of one embodiment of the silent face liveness detection device of the present invention;

[0037] Figure 8 Schematic diagram of a training module of a convolutional neural network of the present invention;

[0038] Figure 9 for Figure 8 a schematic diagram of the image preparation module;

[0039] Figure 10 is a schematic diagram of another embodiment of the silent face liveness detection device of the present invention;

[0040] Figure 11 for Figure 7 a schematic diagram of the processing module;

[0041] Figure 12This is a structural diagram of one embodiment of the convolutional neural network of the present invention. DETAILED DESCRIPTION

[0042] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0043] Example 1:

[0044] The embodiment of the present invention provides a silent liveness detection method for a face, which can effectively distinguish the features of living and non-living objects through automatic learning of a convolutional neural network (CNN), solving the problem that manually designed features rely on human experience and have poor robustness.

[0045] Convolutional neural networks require pre-training before use. During training, each pixel in the training sample is labeled with a value, which indicates whether the pixel is live or prosthetic (or whether the pixel and its surrounding pixels are live or prosthetic). Generally, a label value of 0 indicates prosthetic, and a label value of 1 indicates live.

[0046] When labeling each pixel in a training sample, it is generally done by region. For example, all pixels in a certain region are labeled as living, and all pixels in another region are labeled as prosthetic. Of course, it is also possible to label each pixel individually instead of by region. Whether the pixels are labeled by region or individually, it does not affect the implementation of the present invention.

[0047] After the convolutional neural network training is completed, silent face liveness detection can be performed, such as Figure 1 As shown, the silent face liveness detection method includes:

[0048] S100: Input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is the probability value of the living body or the prosthesis in the receptive field of the corresponding position of the image to be detected.

[0049] The image to be detected is the portion of a face captured by a camera. This image can be either a standard visible light image or a near-infrared image. Visible light imaging eliminates the need for additional 3D structured light sources or near-infrared light sources, and visible light imaging has low requirements for equipment cost and lighting conditions.

[0050] The present invention does not limit the method for obtaining the image to be detected. For example, a face image captured by a camera can be subjected to operations such as face detection, positioning, cropping, and normalization to obtain the image to be detected.

[0051] The image to be detected is fed into a convolutional neural network for processing, generating a target feature map. After processing by the convolutional neural network, each position on the target feature map corresponds to a receptive field of a specific size on the image to be detected.

[0052] Assume that the convolutional neural network is downsampled by 4 times as a whole, and the 4N*4N image to be detected is calculated to obtain the target feature map of N*N size. Figure 5 The correspondence between the position (i, j) on the target feature map and the receptive field on the image to be detected is given in . For the position (i, j) on the target feature map, the receptive field corresponding to the center of the image to be detected is the position (4*i,4*j), and the size of the receptive field is 4*4 pixels.

[0053] Since each pixel of the training sample is annotated with a label value representing a live object or a prosthesis during convolutional neural network training, when the trained convolutional neural network processes the image to be detected, the feature value of each position on the target feature map obtained represents the probability value of a live object or a prosthesis in the corresponding receptive field.

[0054] Assume that the feature value of position (i, j) on the target feature map is s ij , then s ij Represents the probability value of the receptive field corresponding to the (4*i, 4*j) position of the image to be detected as a living body or a prosthesis, which represents the judgment result of the liveness detection of the receptive field area of ​​the image to be detected after the CNN is passed through the receptive field area of ​​the image to be detected.

[0055] That is to say, the feature value of each position on the target feature map represents the probability of a small area on the image to be detected being alive or non-living.

[0056] S200: Statistically analyzing the feature values ​​at all positions on the target feature map to obtain a liveness detection score for the image to be detected.

[0057] The eigenvalue at each location on the target feature map represents the probability of liveness or non-liveness for a small area of ​​the image to be detected. Statistical analysis of the eigenvalues ​​at all locations yields a liveness detection score for the entire image to be detected. Because the liveness detection score for the entire image to be detected is derived from the liveness or non-liveness probabilities of each local area, it fully utilizes local information in the facial image, capturing richer details.

[0058] The present invention does not limit the specific method of statistical analysis. For example, the statistical analysis method is described as a weighted method: weights are set for the feature values ​​at each position and voting is performed. The final weighted score is the final liveness detection score. In one example, the weights of each position are set equal, that is, the final liveness detection score s is expressed as:

[0059]

[0060] S300: Compare the threshold set by the liveness detection score of the image to be detected. If the threshold is met, it is judged to be a living body; otherwise, it is judged to be a prosthesis.

[0061] The present invention does not limit the setting method of the threshold and the method of comparing the liveness detection score with the threshold. Exemplarily, the threshold is set to T, and s is compared with the given threshold T. If s>T, it is judged to be a living body, otherwise it is a prosthesis.

[0062] The present invention uses a convolutional neural network to perform liveness detection on facial images. When training the CNN, each pixel in the training sample is labeled with a live or prosthetic label value. The image to be detected is then input into the trained CNN to obtain a target feature map. The eigenvalues ​​at each position on the target feature map represent the probability value of the corresponding receptive field as being live or prosthetic. After obtaining the target feature map, the present invention no longer uses softmax to directly classify the image into two or more categories as in the prior art CNN liveness detection methods. Instead, the target feature map obtained by the convolutional neural network is operated on, and the eigenvalues ​​at all positions in the target feature map are statistically analyzed to obtain the final liveness detection score. Finally, the final liveness detection result is obtained based on the liveness detection score and the set threshold.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. Use convolutional neural networks for liveness detection, which has good robustness. It can also detect liveness in visible light facial images, which can reduce equipment costs and has low requirements for lighting environments.

[0065] 2. Each pixel in the training sample is marked with a live or prosthetic label value, and the live and prosthetic images are annotated point by point for liveness judgment. This can not only improve the accuracy of liveness judgment, but also effectively reflect which areas have the highest impact on liveness detection.

[0066] 3. Operate the target feature map obtained by CNN on the image to be detected, and judge the liveness judgment result on a smaller receptive field of the image to be detected. No longer use softmax to reduce the possibility of overfitting. At the same time, make full use of the local detail information of the face to make the liveness judgment more reliable.

[0067] As mentioned above, the convolutional neural network needs to be trained before use. The convolutional neural network of the present invention is trained as follows: Figure 2 The method shown is trained to obtain:

[0068] S100': preparing a plurality of real face images and prosthetic face images, dividing the real face images into a plurality of real sub-regions, and dividing the prosthetic face images into a plurality of prosthetic sub-regions corresponding to the positions of the plurality of real sub-regions.

[0069] During the training process, it is first necessary to prepare a training sample set consisting of training samples. The training samples of the present invention are obtained by processing multiple real human face images and prosthetic human face images.

[0070] The present invention does not limit the division method of the real person sub-region and the prosthesis sub-region, as long as the position of the prosthesis sub-region corresponds to the real person sub-region.

[0071] S200': Randomly overlaying a plurality of prosthetic sub-regions onto the real person sub-regions at corresponding positions of the real person face image to obtain a composite image.

[0072] When randomly overlaying the prosthetic sub-regions on the real face image, there is no limit on the number of prosthetic sub-regions, which can be one, multiple, or zero, that is, the real face image is not covered.

[0073] In the composite image, part of it is a prosthesis, part of it is a living body, or both of it are prostheses, or both of it are living bodies.

[0074] S300 ′: labeling the label values ​​of all pixels in the prosthesis sub-region on the synthetic image as values ​​representing the prosthesis, and labeling the label values ​​of all pixels in other regions as values ​​representing the living body, to obtain training samples.

[0075] Generally, the prosthetic face is marked as 0 and the living face is marked as 1. Multiple real face images and prosthetic face images can be used to obtain multiple training samples to form a training sample set.

[0076] S400': inputting the training sample into the convolutional neural network for training, wherein the label value of each position on the target feature map obtained during training is a statistical value of the label values ​​of all pixels within the receptive field of the corresponding position of the training sample.

[0077] This step first initializes the parameters of the convolutional neural network and inputs the prepared training samples into the convolutional neural network for parameter training.

[0078] During training, assuming the convolutional neural network is downsampled by a factor of 4, the 4N*4N training sample is computed to produce an N*N target feature map. The (i, j) position on the target feature map corresponds to a 4*4 pixel receptive field with the training sample centered at (4*i, 4*j).

[0079] During training, the label values ​​are also downsampled by a factor of 4. The label value at position (i, j) in the target feature map is the statistical value of the label values ​​of all pixels within the 4*4 receptive field of the training sample position (4*i, 4*j). Generally, this statistical value is the average value, that is, the label value at each position in the target feature map is the average of the labels of the corresponding receptive field of the training sample.

[0080] The present invention does not limit the specific method of dividing the real person sub-region and the prosthetic sub-region. For example, Figure 3 As shown, the aforementioned S100' includes:

[0081] S110': performing face detection and key point positioning on the real face image to obtain the real face region and real face key points; performing face detection and key point positioning on the prosthetic face image to obtain the prosthetic face region and prosthetic face key points.

[0082] This step is used to obtain the position and key point coordinates of the real face area on the real face image, as well as the position and key point coordinates of the prosthetic face area on the prosthetic face image, which are used as the positioning reference for the subsequent division of real sub-areas and prosthetic sub-areas.

[0083] S120': Divide the real face region into a plurality of real sub-regions according to the real face key points, and divide the prosthetic face region into a plurality of prosthetic sub-regions corresponding to the positions of the plurality of real sub-regions according to the prosthetic face key points.

[0084] For example, Figure 4 As shown in the figure, assuming that 106 key points are located on the real / prosthetic face image, the key points of the real face area are divided into the real left eye area, the real right eye area, the real entire eye area, the real left eyebrow area, the real right eyebrow area, the real entire eyebrow area, the real nose area, the real left cheek area, the real right cheek area, the real chin area, the real forehead area and the real face contour area;

[0085] Correspondingly, the key points of the prosthetic face area are divided into the prosthetic left eye area, the prosthetic right eye area, the prosthetic entire eye area, the prosthetic left eyebrow area, the prosthetic right eyebrow area, the prosthetic entire eyebrow area, the prosthetic nose area, the prosthetic left cheek area, the prosthetic right cheek area, the prosthetic chin area, the prosthetic forehead area and the prosthetic face contour area.

[0086] Then, cut and paste the image of a certain prosthetic sub-region on the prosthetic face image to the corresponding real sub-region on the real face image, and annotate the pixels on the synthesized image one by one, for example:

[0087] Based on the coordinates of the located key points, the image of the prosthetic left eye area on the prosthetic face image is synthesized onto the real left eye area on the real face image. The pixels in the left eye area of ​​the synthesized image are labeled as 0, and the rest of the image is labeled as 1, completing the labeling of a training sample. For example, the key points in the synthesized areas numbered 52, 53, 54, 55, 56, 57, 58, 59, 60, and 61 are labeled as 0, and the rest of the area is labeled as 1.

[0088] Similarly, each subregion of the real face images and each subregion of the prosthetic face images were cropped and synthesized, and the synthesized images were annotated pixel by pixel. This means that each position in the 4N*4N image was labeled with a 1 if the position was a live cell and a 0 if it was a prosthetic cell, forming a training sample set.

[0089] When training CNN, the present invention divides the training samples into regions and labels each pixel, expands the range of samples, and divides the smaller pixel areas of the face into living and prosthetic. In this way, the local information of the face image can be fully utilized when training the convolutional neural network to obtain richer details.

[0090] The present invention does not limit the method for obtaining the image to be detected. In one example, Figure 6 As shown, before S100, it also included:

[0091] S10: Acquire a face image and perform face detection and key point positioning to obtain a face area.

[0092] The face image is an image captured by a camera. After face detection and key point positioning, the facial area position and the coordinates of the key points are obtained. The present invention does not limit the face detection and key point positioning methods. The face detection method can be MTCNN, Harr-AdaBoost, SSD, Faster RCNN, CenterNet and other algorithms. The feature point positioning method can also be SDM, LBF, LAB and other algorithms.

[0093] After obtaining the facial region, a preliminary judgment of whether the subject is alive can be made based on the skin color of the facial region. The judgment method includes:

[0094] S20: Determine whether the color of the facial area meets the preset skin color requirement. If so, scale the facial area to obtain an image to be detected. Otherwise, determine that the facial area is a prosthesis.

[0095] This step first performs skin color detection to determine whether the facial area meets the skin color requirements. This is used to quickly resolve liveness detection tasks for non-normal facial skin colors, such as grayscale images. This determination typically involves presetting a skin color threshold range and then comparing the color of the facial area to the threshold. If the color falls within the threshold range, the pre-set skin color requirement is considered met. Otherwise, it is considered not to meet the pre-set skin color requirement, and a prosthesis judgment is directly output, ending the process.

[0096] If the preset skin color requirements are met, the facial area image needs to be further processed (such as normalization) to meet the input requirements of the convolutional neural network.

[0097] The processing method of the present invention is image scaling. For example, the facial area is scaled to an image to be detected with a width and height of 4N pixels using a linear interpolation method. It should be noted that the facial area in the present invention only needs to be scaled in width and height, and no face alignment operation is required.

[0098] The image to be detected is a three-channel RGB image. After obtaining the image to be detected, it can be input into the convolutional neural network for liveness detection. The structure of the convolutional neural network is as follows: Figure 12 As shown, Figure 12 In the figure, the dotted line is the dividing line. It is not the main structure of the convolutional neural network, but is only used to segment the diagram. The numbers on the left and right sides represent the network connection labels, and lines with the same label are connected.

[0099] Figure 12 The convolutional neural network with the structure shown processes the image to be detected as follows:

[0100] S101: The image to be detected (Input) undergoes a first operation (Process1) to obtain a first feature map (Feature1), the first feature map (Feature1) undergoes a second operation (Process2) to obtain a second feature map (Feature2), and the first feature map (Feature1) and the second feature map (Feature2) undergo a first Eltwise operation (Eltwise1) to obtain a third feature map (Feature3).

[0101] S102: The third feature map (Feature3) is subjected to the third operation (Process3) to obtain the fourth feature map (Feature4), the fourth feature map (Feature4) is subjected to the fourth operation (Process4) to obtain the fifth feature map (Feature5), and the fourth feature map (Feature4) and the fifth feature map (Feature5) are subjected to the second Eltwise operation (Eltwise2) to obtain the sixth feature map (Feature6).

[0102] S103: The sixth feature map (Feature6) is subjected to the fifth operation (Process5) to obtain the seventh feature map (Feature7), and the sixth feature map (Feature6) and the seventh feature map (Feature7) are subjected to the third Eltwise operation (Eltwise3) to obtain the eighth feature map (Feature8).

[0103] S104: The eighth feature map (Feature8) undergoes the sixth operation (Process6) to obtain the ninth feature map (Feature9), the ninth feature map (Feature9) undergoes the seventh operation (Process7) to obtain the tenth feature map (Feature10), the tenth feature map (Feature10) and the ninth feature map (Feature9) undergo the fourth Eltwise operation (Eltwise4) to obtain the eleventh feature map (Feature11).

[0104] S105: The eleventh feature map (Feature11) is processed through the eighth process (Process8) to obtain the twelfth feature map (Feature12), and the eleventh feature map (Feature11) and the twelfth feature map (Feature12) are processed through the fifth Eltwise process (Eltwise5) to obtain the thirteenth feature map (Feature13).

[0105] S106: The thirteenth feature map (Feature13) is subjected to the ninth operation (Process9) to obtain the fourteenth feature map (Feature14), and the thirteenth feature map (Feature13) and the fourteenth feature map (Feature14) are subjected to the sixth Eltwise operation (Eltwise6) to obtain the fifteenth feature map (Feature15).

[0106] S107: The fifteenth feature map (Feature15) undergoes the tenth operation (Process10) to obtain the sixteenth feature map (Feature16), the sixteenth feature map (Feature16) undergoes the eleventh operation (Process11) to obtain the seventeenth feature map (Feature17), the sixteenth feature map (Feature16) and the seventeenth feature map (Feature17) undergo the seventh Eltwise operation (Eltwise7) to obtain the eighteenth feature map (Feature18).

[0107] S108: The eighteenth feature map (Feature18) is processed through the twelfth process (Process12) to obtain the nineteenth feature map (Feature19), and the eighteenth feature map (Feature18) and the nineteenth feature map (Feature19) are processed through the eighth Eltwise process (Eltwise8) to obtain the twentieth feature map (Feature20).

[0108] S109: The twentieth feature map (Feature20) undergoes the thirteenth operation (Process13) to obtain the twenty-first feature map (Feature21), the twenty-first feature map (Feature21) undergoes the fourteenth operation (Process14) to obtain the twenty-second feature map (Feature22), the twenty-first feature map (Feature21) and the twenty-second feature map (Feature22) undergo the ninth Eltwise operation (Eltwise9) to obtain the twenty-third feature map (Feature23).

[0109] S110: The twenty-third feature map (Feature23) is subjected to the fifteenth process (Process15) to obtain the twenty-fourth feature map (Feature24), and the twenty-third feature map (Feature23) and the twenty-fourth feature map (Feature23) are subjected to the tenth Eltwise process (Eltwise10) to obtain the twenty-fifth feature map (Feature25).

[0110] S111: The twenty-fifth feature map (Feature25) undergoes the sixteenth operation (Process16) to obtain the twenty-sixth feature map (Feature26), the twentieth feature map (Feature20) undergoes the seventeenth operation (Process17) to obtain the twenty-seventh feature map (Feature27), and the twenty-sixth feature map (Feature26) and the twenty-seventh feature map (Feature27) undergo the eleventh Eltwise operation (Eltwise11) to obtain the twenty-eighth feature map (Feature28).

[0111] S112: The twenty-eighth feature map (Feature28) undergoes the eighteenth operation (Process18) to obtain the twenty-ninth feature map (Feature29), the eighth feature map (Feature8) undergoes the nineteenth operation (Process19) to obtain the thirtieth feature map (Feature30), and the twenty-ninth feature map (Feature29) and the thirtieth feature map (Feature30) undergo the twelfth Eltwise operation (Eltwise12) to obtain the thirty-first feature map (Feature31).

[0112] S113: The thirty-first feature map (Feature31) undergoes the twentieth operation (Process20) to obtain the thirty-second feature map (Feature32), the third feature map (Feature3) undergoes the twenty-first operation (Process21) to obtain the thirty-third feature map (Feature33), and the thirty-second feature map (Feature32) and the thirty-third feature map (Feature33) undergo the thirteenth Eltwise operation (Eltwise13) to obtain the thirty-fourth feature map (Feature34).

[0113] S114: The thirty-fourth feature map (Feature34) is subjected to the twenty-second operation (Process22) to obtain the target feature map (Output).

[0114] Among them, the first operation to the twenty-second operation all include a convolution operation, a BN operation, and an activation operation.

[0115] Assume that the aforementioned convolutional neural network performs a 4x downsampling. After computing the 4N*4N image to be detected, an N*N target feature map is generated. The feature value at each position on the target feature map represents the probability of judging the corresponding receptive field as live or prosthetic. For example, the image to be detected is 80*80 pixels in size, and the target feature map is 20*20 pixels in size.

[0116] The present invention synthesizes the prosthetic face image and the real face image by dividing them into regions, and annotates the synthesized image pixel by pixel, so that the local information of the face image can be fully utilized when training the convolutional neural network to obtain richer details. In addition, the feature map obtained by CNN is operated to determine the original Figure 1 Liveness detection results based on a smaller receptive field are no longer based on softmax, reducing the possibility of overfitting and making classification more reliable while fully utilizing facial information. Furthermore, point-by-point annotation of live and prosthetic images not only facilitates liveness determination but also effectively identifies which areas have the highest impact on liveness detection.

[0117] Example 2:

[0118] The embodiment of the present invention provides a silent face liveness detection device, such as Figure 7 As shown, the device includes:

[0119] The processing module 100 is used to input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is the probability value of the living body or the prosthesis in the receptive field of the corresponding position of the image to be detected.

[0120] The statistical module 200 is used to perform statistical analysis on the feature values ​​of all positions on the target feature map to obtain a liveness detection score of the image to be detected.

[0121] The judgment module 300 is used to compare the threshold set by the liveness detection score of the image to be detected. If the threshold is met, it is judged as a living body; otherwise, it is judged as a prosthesis.

[0122] When the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing a living body or a prosthesis.

[0123] The convolutional neural network of the present invention is Figure 8 The module shown is trained to produce:

[0124] The image preparation module 100' is used to prepare multiple real face images and prosthetic face images, and divide the real face images into multiple real sub-regions, and divide the prosthetic face images into multiple prosthetic sub-regions corresponding to the positions of the multiple real sub-regions.

[0125] The image synthesis module 200 ′ is used to randomly overlay a plurality of prosthetic sub-regions onto the real human sub-regions at corresponding positions of the real human face image to obtain a synthesized image.

[0126] The labeling module 300' is used to label the label values ​​of all pixels in the prosthesis sub-region on the synthetic image as values ​​representing the prosthesis, and label the label values ​​of all pixels in other regions as values ​​representing the living body, to obtain training samples.

[0127] The training module 400' is used to input the training samples into the convolutional neural network for training, wherein the label value of each position on the target feature map obtained during training is the statistical value of the label values ​​of all pixels in the receptive field of the corresponding position of the training sample.

[0128] like Figure 9 As shown, the aforementioned image preparation module 100' includes:

[0129] The detection and positioning unit 110' is used to perform face detection and key point positioning on real face images to obtain real face areas and real face key points, and to perform face detection and key point positioning on prosthetic face images to obtain prosthetic face areas and prosthetic face key points.

[0130] The division unit 120' is configured to divide the real face region into a plurality of real sub-regions according to the real face key points, and to divide the prosthetic face region into a plurality of prosthetic sub-regions corresponding to the positions of the plurality of real sub-regions according to the prosthetic face key points.

[0131] Among them, the real person sub-areas include the real person left eye area, the real person right eye area, the real person entire eye area, the real person left eyebrow area, the real person right eyebrow area, the real person entire eyebrow area, the real person nose area, the real person left cheek area, the real person right cheek area, the real person chin area, the real person forehead area and the real person face contour area.

[0132] The prosthetic sub-areas include the prosthetic left eye area, the prosthetic right eye area, the prosthetic entire eye area, the prosthetic left eyebrow area, the prosthetic right eyebrow area, the prosthetic entire eyebrow area, the prosthetic nose area, the prosthetic left cheek area, the prosthetic right cheek area, the prosthetic chin area, the prosthetic forehead area and the prosthetic facial contour area.

[0133] The device of the present invention also includes Figure 10 The modules shown are:

[0134] The acquisition module 10 is used to acquire a face image and perform face detection and key point positioning to obtain a face area.

[0135] The skin color judgment module 20 is used to judge whether the color of the facial area meets the preset skin color requirement. If so, the facial area is scaled to obtain the image to be detected. Otherwise, it is judged as a prosthesis.

[0136] As a specific example, Figure 11 As shown, the aforementioned processing module 100 includes:

[0137] The first operation unit 101 is configured to perform a first operation on the image to be detected to obtain a first feature map, perform a second operation on the first feature map to obtain a second feature map, and perform a first Eltwise operation on the first feature map and the second feature map to obtain a third feature map.

[0138] The second operation unit 102 is configured to perform a third operation on the third feature map to obtain a fourth feature map, perform a fourth operation on the fourth feature map to obtain a fifth feature map, and perform a second Eltwise operation on the fourth feature map and the fifth feature map to obtain a sixth feature map.

[0139] The third operation unit 103 is configured to perform the fifth operation on the sixth feature map to obtain a seventh feature map, and perform the third Eltwise operation on the sixth feature map and the seventh feature map to obtain an eighth feature map.

[0140] The fourth operation unit 104 is configured to perform the sixth operation on the eighth feature map to obtain a ninth feature map, perform the seventh operation on the ninth feature map to obtain a tenth feature map, and perform the fourth Eltwise operation on the tenth feature map and the ninth feature map to obtain an eleventh feature map.

[0141] The fifth operation unit 105 is configured to perform the eighth operation on the eleventh feature map to obtain a twelfth feature map, and perform the fifth Eltwise operation on the eleventh feature map and the twelfth feature map to obtain a thirteenth feature map.

[0142] The sixth operation unit 106 is configured to perform the ninth operation on the thirteenth feature map to obtain a fourteenth feature map, and perform the sixth Eltwise operation on the thirteenth feature map and the fourteenth feature map to obtain a fifteenth feature map.

[0143] The seventh operation unit 107 is configured to perform a tenth operation on the fifteenth feature map to obtain a sixteenth feature map, perform an eleventh operation on the sixteenth feature map to obtain a seventeenth feature map, and perform a seventh Eltwise operation on the sixteenth feature map and the seventeenth feature map to obtain an eighteenth feature map.

[0144] The eighth operation unit 108 is configured to perform the twelfth operation on the eighteenth feature map to obtain a nineteenth feature map, and perform the eighth Eltwise operation on the eighteenth feature map and the nineteenth feature map to obtain a twentieth feature map.

[0145] The ninth operation unit 109 is configured to perform the thirteenth operation on the twentieth feature map to obtain a twenty-first feature map, perform the fourteenth operation on the twenty-first feature map to obtain a twenty-second feature map, and perform the ninth Eltwise operation on the twenty-first and twenty-second feature maps to obtain a twenty-third feature map.

[0146] The tenth operation unit 110 is configured to perform the fifteenth operation on the twenty-third feature map to obtain a twenty-fourth feature map, and perform the tenth Eltwise operation on the twenty-third feature map and the twenty-fourth feature map to obtain a twenty-fifth feature map.

[0147] The eleventh operation unit 111 is configured to perform the sixteenth operation on the twenty-fifth feature map to obtain a twenty-sixth feature map, perform the seventeenth operation on the twentieth feature map to obtain a twenty-seventh feature map, and perform the eleventh Eltwise operation on the twenty-sixth feature map and the twenty-seventh feature map to obtain a twenty-eighth feature map.

[0148] The twelfth operation unit 112 is used to perform the eighteenth operation on the twenty-eighth feature map to obtain the twenty-ninth feature map, perform the nineteenth operation on the eighth feature map to obtain the thirtieth feature map, and perform the twelfth Eltwise operation on the twenty-ninth feature map and the thirtieth feature map to obtain the thirty-first feature map.

[0149] The thirteenth operation unit 113 is used to perform the twentieth operation on the thirty-first feature map to obtain a thirty-second feature map, perform the twenty-first operation on the third feature map to obtain a thirty-third feature map, and perform the thirteenth Eltwise operation on the thirty-second feature map and the thirty-third feature map to obtain a thirty-fourth feature map.

[0150] The fourteenth operation unit 114 is configured to obtain a target feature map by performing the twenty-second operation on the thirty-fourth feature map.

[0151] The aforementioned first to twenty-second operations all include a convolution operation, a BN operation, and an activation operation.

[0152] The present invention synthesizes the prosthetic face image and the real face image by dividing them into regions, and annotates the synthesized image pixel by pixel, so that the local information of the face image can be fully utilized when training the convolutional neural network to obtain richer details. In addition, the feature map obtained by CNN is operated to determine the original Figure 1 Liveness detection results based on a smaller receptive field are no longer based on softmax, reducing the possibility of overfitting and making classification more reliable while fully utilizing facial information. Furthermore, point-by-point annotation of live and prosthetic images not only facilitates liveness determination but also effectively identifies which areas have the highest impact on liveness detection.

[0153] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment 1. For the sake of brevity, any matters not mentioned in the device embodiment can be referred to the corresponding contents in the aforementioned method embodiment 1. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working processes of the devices and units described above can all refer to the corresponding processes in the aforementioned method embodiment 1 and will not be repeated here.

[0154] Example 3:

[0155] The method described in Example 1 above provided by the present invention can implement the business logic through a computer program and record it on a storage medium. The storage medium can be read and executed by a computer to achieve the effects of the solution described in Example 1 of this specification. Therefore, the present invention also provides a computer-readable storage medium for silent face liveness detection, including a memory for storing processor-executable instructions. When the instructions are executed by the processor, the steps of the silent face liveness detection method in Example 1 are implemented.

[0156] The storage medium may include a physical device for storing information, typically digitizing the information and then storing it in a medium utilizing electrical, magnetic, or optical means. Examples of such storage media include: devices that use electrical energy to store information, such as various types of memory, such as RAM and ROM; devices that use magnetic energy to store information, such as hard disks, floppy disks, magnetic tapes, magnetic core memories, bubble memories, and USB flash drives; and devices that use optical means to store information, such as CDs or DVDs. Of course, there are also other types of readable storage media, such as quantum memories and graphene memories.

[0157] The present invention synthesizes the prosthetic face image and the real face image by dividing them into regions, and annotates the synthesized image pixel by pixel, so that the local information of the face image can be fully utilized when training the convolutional neural network to obtain richer details. In addition, the feature map obtained by CNN is operated to determine the original Figure 1 Liveness detection results based on a smaller receptive field are no longer based on softmax, reducing the possibility of overfitting and making classification more reliable while fully utilizing facial information. Furthermore, point-by-point annotation of live and prosthetic images not only facilitates liveness determination but also effectively identifies which areas have the highest impact on liveness detection.

[0158] The storage medium described above may also include other implementation methods according to the description of method embodiment 1. The implementation principle and technical effects produced by this embodiment are the same as those of the aforementioned method embodiment 1. For details, please refer to the description of the relevant method embodiment 1 and will not be repeated here.

[0159] Example 4:

[0160] The present invention also provides a device for silent face liveness detection. The device may be a standalone computer or may include an actual operating device that uses one or more methods or one or more embodiments of the apparatus described in this specification. The device may include at least one processor and a memory storing computer-executable instructions. When the processor executes the instructions, it implements the steps of the silent face liveness detection method described in any one or more of Embodiment 1.

[0161] The present invention synthesizes the prosthetic face image and the real face image by dividing them into regions, and annotates the synthesized image pixel by pixel, so that the local information of the face image can be fully utilized when training the convolutional neural network to obtain richer details. In addition, the feature map obtained by CNN is operated to determine the original Figure 1Liveness detection results based on a smaller receptive field are no longer based on softmax, reducing the possibility of overfitting and making classification more reliable while fully utilizing facial information. Furthermore, point-by-point annotation of live and prosthetic images not only facilitates liveness determination but also effectively identifies which areas have the highest impact on liveness detection.

[0162] The above-mentioned device can also include other implementation methods according to the description of method embodiment 1. The implementation principle and technical effects produced by this embodiment are the same as those of the aforementioned method embodiment 1. For details, please refer to the description of the relevant method embodiment 1 and will not be repeated here.

[0163] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, intended to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can, within the technical scope disclosed by the present invention, modify or readily conceive of variations to the technical solutions described in the above-described embodiments, or substitute equivalently for some of the technical features thereof. Such modifications, variations, or substitutions do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. They should all be encompassed within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for silent face liveness detection, characterized in that: The method comprises: Input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is the probability value of the receptive field of the corresponding position of the image to be detected being a living object or a prosthesis; Perform statistical analysis on the feature values ​​of all positions on the target feature map to obtain the liveness detection score of the image to be detected; Compare the liveness detection score of the image to be detected with the threshold set. If it meets the set threshold, it is judged as a live body, otherwise it is judged as a fake body; When the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing whether the pixel is a living object or a prosthesis; or each pixel of the training sample is annotated with a label value representing whether the pixel and its surrounding pixels are a living object or a prosthesis; The convolutional neural network is trained by the following method: Preparing multiple real face images and prosthetic face images, dividing the real face images into multiple real sub-regions, and dividing the prosthetic face images into multiple prosthetic sub-regions corresponding to the positions of the multiple real sub-regions; Randomly overlay several prosthetic sub-regions onto the real sub-regions at the corresponding positions of the real face image to obtain a synthetic image; The label values ​​of all pixels in the prosthesis sub-region on the synthetic image are marked as the values ​​representing the prosthesis, and the label values ​​of all pixels in other regions are marked as the values ​​representing the living body, thereby obtaining training samples; The training samples are input into a convolutional neural network for training, wherein the label value of each position on the target feature map obtained during training is a statistical value of the label values ​​of all pixels within the receptive field of the corresponding position of the training sample.

2. The method for silent face liveness detection according to claim 1, wherein: The method of dividing the real face image into a plurality of real sub-regions and dividing the prosthetic face image into a plurality of prosthetic sub-regions corresponding to the positions of the plurality of real sub-regions comprises: Perform face detection and key point positioning on real face images to obtain real face regions and real face key points; perform face detection and key point positioning on prosthetic face images to obtain prosthetic face regions and prosthetic face key points; Dividing the real face area into a plurality of real sub-areas according to the key points of the real face, and dividing the prosthetic face area into a plurality of prosthetic sub-areas corresponding to the positions of the plurality of real sub-areas according to the key points of the prosthetic face; The real person sub-regions include the real person left eye region, the real person right eye region, the real person entire eye region, the real person left eyebrow region, the real person right eyebrow region, the real person entire eyebrow region, the real person nose region, the real person left cheek region, the real person right cheek region, the real person chin region, the real person forehead region and the real person face contour region; The prosthesis sub-areas include the prosthesis left eye area, the prosthesis right eye area, the prosthesis entire eye area, the prosthesis left eyebrow area, the prosthesis right eyebrow area, the prosthesis entire eyebrow area, the prosthesis nose area, the prosthesis left cheek area, the prosthesis right cheek area, the prosthesis chin area, the prosthesis forehead area and the prosthesis facial contour area.

3. The method for silent face liveness detection according to claim 1, wherein: The image to be detected is input into a pre-trained convolutional neural network to obtain a target feature map, which also includes: Obtain a face image and perform face detection and key point positioning to obtain the face area; Determine whether the color of the facial area meets the preset skin color requirement. If so, scale the facial area to obtain the image to be detected. Otherwise, determine it as a prosthesis.

4. The method for silent face liveness detection according to any one of claims 1 to 3, characterized in that: The image to be detected is input into a pre-trained convolutional neural network to obtain a target feature map, including: The image to be detected is subjected to a first operation to obtain a first feature map, the first feature map is subjected to a second operation to obtain a second feature map, and the first feature map and the second feature map are subjected to a first Eltwise operation to obtain a third feature map; The third feature map is subjected to a third operation to obtain a fourth feature map, the fourth feature map is subjected to a fourth operation to obtain a fifth feature map, and the fourth feature map and the fifth feature map are subjected to a second Eltwise operation to obtain a sixth feature map; The sixth feature map is subjected to the fifth operation to obtain a seventh feature map, and the sixth feature map and the seventh feature map are subjected to the third Eltwise operation to obtain an eighth feature map; The eighth feature map is subjected to the sixth operation to obtain a ninth feature map, the ninth feature map is subjected to the seventh operation to obtain a tenth feature map, and the tenth feature map and the ninth feature map are subjected to the fourth Eltwise operation to obtain an eleventh feature map; The eleventh feature map is subjected to the eighth operation to obtain a twelfth feature map, and the eleventh feature map and the twelfth feature map are subjected to the fifth Eltwise operation to obtain a thirteenth feature map; The thirteenth feature map is subjected to the ninth operation to obtain a fourteenth feature map, and the thirteenth feature map and the fourteenth feature map are subjected to the sixth Eltwise operation to obtain a fifteenth feature map; The fifteenth feature map is subjected to a tenth operation to obtain a sixteenth feature map, the sixteenth feature map is subjected to an eleventh operation to obtain a seventeenth feature map, and the sixteenth feature map and the seventeenth feature map are subjected to a seventh Eltwise operation to obtain an eighteenth feature map; The eighteenth feature map is subjected to the twelfth operation to obtain a nineteenth feature map, and the eighteenth feature map and the nineteenth feature map are subjected to the eighth Eltwise operation to obtain a twentieth feature map; The 20th feature map is subjected to the 13th operation to obtain a 21st feature map, the 21st feature map is subjected to the 14th operation to obtain a 22nd feature map, and the 21st feature map and the 22nd feature map are subjected to the 9th Eltwise operation to obtain a 23rd feature map; The 23rd feature map is subjected to the 15th operation to obtain the 24th feature map, and the 23rd feature map and the 24th feature map are subjected to the 10th Eltwise operation to obtain the 25th feature map; The 25th feature map is subjected to the 16th operation to obtain the 26th feature map, the 20th feature map is subjected to the 17th operation to obtain the 27th feature map, and the 26th feature map and the 27th feature map are subjected to the 11th Eltwise operation to obtain the 28th feature map; The 28th feature map is subjected to the 18th operation to obtain the 29th feature map, the 8th feature map is subjected to the 19th operation to obtain the 30th feature map, and the 29th feature map and the 30th feature map are subjected to the 12th Eltwise operation to obtain the 31st feature map; The thirty-first feature map is subjected to the twentieth operation to obtain a thirty-second feature map, the third feature map is subjected to the twenty-first operation to obtain a thirty-third feature map, and the thirty-second feature map and the thirty-third feature map are subjected to the thirteenth Eltwise operation to obtain a thirty-fourth feature map; The target feature map is obtained by performing the twenty-second operation on the thirty-fourth feature map; The first to twenty-second operations all include a convolution operation, a BN operation, and an activation operation.

5. A silent face liveness detection device, characterized in that: The device comprises: A processing module is configured to input the image to be detected into a pre-trained convolutional neural network to obtain a target feature map, wherein the feature value of each position on the target feature map is a probability value of a living object or a prosthesis in the receptive field of the corresponding position of the image to be detected; The statistical module is used to perform statistical analysis on the feature values ​​of all positions on the target feature map to obtain the liveness detection score of the image to be detected; A judgment module is used to compare the liveness detection score of the image to be detected with a set threshold. If the set threshold is met, it is judged to be alive, otherwise it is judged to be a prosthesis; When the convolutional neural network is trained, each pixel of the training sample is annotated with a label value representing whether the pixel is a living object or a prosthesis; or each pixel of the training sample is annotated with a label value representing whether the pixel and its surrounding pixels are a living object or a prosthesis; The convolutional neural network is trained by the following modules: An image preparation module, configured to prepare a plurality of real face images and prosthetic face images, and to divide the real face images into a plurality of real sub-regions, and to divide the prosthetic face images into a plurality of prosthetic sub-regions corresponding to the positions of the plurality of real sub-regions; An image synthesis module is used to randomly overlay several prosthetic sub-regions onto the real sub-regions at corresponding positions in the real face image to obtain a synthesized image; a labeling module for labeling the label values ​​of all pixels in the prosthesis sub-region on the synthetic image as values ​​representing the prosthesis, and labeling the label values ​​of all pixels in other regions as values ​​representing the living body, thereby obtaining training samples; A training module is used to input the training samples into a convolutional neural network for training, wherein the label value of each position on the target feature map obtained during training is a statistical value of the label values ​​of all pixels within the receptive field of the corresponding position of the training sample.

6. The silent face liveness detection device according to claim 5, characterized in that: The processing module includes: A first operation unit is configured to obtain a first feature map by performing a first operation on the image to be detected, obtain a second feature map by performing a second operation on the first feature map, and obtain a third feature map by performing a first Eltwise operation on the first feature map and the second feature map; a second operation unit, configured to perform a third operation on the third feature map to obtain a fourth feature map, perform a fourth operation on the fourth feature map to obtain a fifth feature map, and perform a second Eltwise operation on the fourth feature map and the fifth feature map to obtain a sixth feature map; a third operation unit, configured to perform a fifth operation on the sixth feature map to obtain a seventh feature map, and perform a third Eltwise operation on the sixth feature map and the seventh feature map to obtain an eighth feature map; a fourth operation unit, configured to perform a sixth operation on the eighth feature map to obtain a ninth feature map, perform a seventh operation on the ninth feature map to obtain a tenth feature map, and perform a fourth Eltwise operation on the tenth feature map and the ninth feature map to obtain an eleventh feature map; a fifth operation unit, configured to perform an eighth operation on the eleventh feature map to obtain a twelfth feature map, and perform a fifth Eltwise operation on the eleventh feature map and the twelfth feature map to obtain a thirteenth feature map; A sixth operation unit, configured to perform a ninth operation on the thirteenth feature map to obtain a fourteenth feature map, and perform a sixth Eltwise operation on the thirteenth feature map and the fourteenth feature map to obtain a fifteenth feature map; a seventh operation unit, configured to perform a tenth operation on the fifteenth feature map to obtain a sixteenth feature map, perform an eleventh operation on the sixteenth feature map to obtain a seventeenth feature map, and perform a seventh Eltwise operation on the sixteenth and seventeenth feature maps to obtain an eighteenth feature map; An eighth operation unit, configured to perform a twelfth operation on the eighteenth feature map to obtain a nineteenth feature map, and perform an eighth Eltwise operation on the eighteenth feature map and the nineteenth feature map to obtain a twentieth feature map; a ninth operation unit, configured to perform a thirteenth operation on the twentieth feature map to obtain a twenty-first feature map, perform a fourteenth operation on the twenty-first feature map to obtain a twenty-second feature map, and perform a ninth Eltwise operation on the twenty-first feature map and the second second feature map to obtain a twenty-third feature map; A tenth operation unit, configured to perform a fifteenth operation on the twenty-third feature map to obtain a twenty-fourth feature map, and perform a tenth Eltwise operation on the twenty-third feature map and the twenty-fourth feature map to obtain a twenty-fifth feature map; an eleventh operation unit, configured to perform a sixteenth operation on the twenty-fifth feature map to obtain a twenty-sixth feature map, perform a seventeenth operation on the twentieth feature map to obtain a twenty-seventh feature map, and perform an eleventh Eltwise operation on the twenty-sixth feature map and the twenty-seventh feature map to obtain a twenty-eighth feature map; a twelfth operation unit, configured to perform an eighteenth operation on the twenty-eighth feature map to obtain a twenty-ninth feature map, perform a nineteenth operation on the eighth feature map to obtain a thirtieth feature map, and perform a twelfth Eltwise operation on the twenty-ninth feature map and the thirtieth feature map to obtain a thirty-first feature map; a thirteenth operation unit, configured to perform a twentieth operation on the thirty-first feature map to obtain a thirty-second feature map, perform a twenty-first operation on the third feature map to obtain a thirty-third feature map, and perform a thirteenth Eltwise operation on the thirty-second and thirty-third feature maps to obtain a thirty-fourth feature map; a fourteenth operation unit, configured to obtain the target feature map by performing a twenty-second operation on the thirty-fourth feature map; The first to twenty-second operations all include a convolution operation, a BN operation, and an activation operation.

7. A computer-readable storage medium for silent face liveness detection, characterized in that: It includes a memory for storing processor-executable instructions, and when the instructions are executed by the processor, the steps of the silent face liveness detection method according to any one of claims 1 to 4 are implemented.

8. A device for silent face liveness detection, characterized in that: The method comprises at least one processor and a memory storing computer-executable instructions, wherein when the processor executes the instructions, the steps of the method for silent face liveness detection according to any one of claims 1 to 4 are implemented.

Images